Re: [PATCH 2/2] xfs: xfs_reflink_convert_cow() memory allocation deadlock

[Dave Chinner <david@fromorbit.com>]

On Thu, Jun 07, 2018 at 07:46:31AM -0700, Darrick J. Wong wrote:
> On Wed, Jun 06, 2018 at 10:56:50PM -0700, Allison Henderson wrote:
> > On 06/06/2018 10:21 PM, Dave Chinner wrote:
> > > From: Dave Chinner <dchinner@redhat.com>
> > > 
> > > xfs_reflink_convert_cow() manipulates the incore extent list
> > > in GFP_KERNEL context in the IO submission path whilst holding
> > > locked pages under writeback. This is a memory reclaim deadlock
> > > vector. This code is not in a transaction, so any memory allocations
> > > it makes aren't protected via the memalloc_nofs_save() context that
> > > transactions carry.
> > > 
> > > Hence we need to run this call under memalloc_nofs_save() context to
> > > prevent potential memory allocations from being run as GFP_KERNEL
> > > and deadlocking.
> > > 
> > > Signed-Off-By: Dave Chinner <dchinner@redhat.com>
> > > ---
> > >   fs/xfs/xfs_aops.c  | 11 +++++++++++
> > >   fs/xfs/xfs_buf.c   |  1 -
> > >   fs/xfs/xfs_linux.h |  1 +
> > >   3 files changed, 12 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
> > > index 767d53222f31..1eb625fdcb1e 100644
> > > --- a/fs/xfs/xfs_aops.c
> > > +++ b/fs/xfs/xfs_aops.c
> > > @@ -531,8 +531,19 @@ xfs_submit_ioend(
> > >   {
> > >   	/* Convert CoW extents to regular */
> > >   	if (!status && ioend->io_type == XFS_IO_COW) {
> > > +		/*
> > > +		 * Yuk. This can do memory allocation, but is not a
> > > +		 * transactional operation so everything is done in GFP_KERNEL
> > > +		 * context. That can deadlock, because we hold pages in
> > > +		 * writeback state and GFP_KERNEL allocations can block on them.
> > > +		 * Hence we must operate in nofs conditions here.
> > > +		 */
> > > +		unsigned nofs_flag;
> > > +
> > > +		nofs_flag = memalloc_nofs_save();
> > >   		status = xfs_reflink_convert_cow(XFS_I(ioend->io_inode),
> > >   				ioend->io_offset, ioend->io_size);
> > > +		memalloc_nofs_restore(nofs_flag);
> 
> DOH. :)
> 
> > >   	}
> > >   	/* Reserve log space if we might write beyond the on-disk inode size. */
> > > diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
> > > index 980bc48979e9..e9c058e3761c 100644
> > > --- a/fs/xfs/xfs_buf.c
> > > +++ b/fs/xfs/xfs_buf.c
> > > @@ -21,7 +21,6 @@
> > >   #include <linux/migrate.h>
> > >   #include <linux/backing-dev.h>
> > >   #include <linux/freezer.h>
> > > -#include <linux/sched/mm.h>
> > >   #include "xfs_format.h"
> > >   #include "xfs_log_format.h"
> > > diff --git a/fs/xfs/xfs_linux.h b/fs/xfs/xfs_linux.h
> > > index ae1e66fa3f61..1631cf4546f2 100644
> > > --- a/fs/xfs/xfs_linux.h
> > > +++ b/fs/xfs/xfs_linux.h
> > > @@ -26,6 +26,7 @@ typedef __u32			xfs_nlink_t;
> > >   #include <linux/semaphore.h>
> > >   #include <linux/mm.h>
> > > +#include <linux/sched/mm.h>
> > >   #include <linux/kernel.h>
> > >   #include <linux/blkdev.h>
> > >   #include <linux/slab.h>
> > > 
> > Looks, ok.  Was moving the header include intentional?  Just clean up maybe?
> > Other than that, looks good.
> 
> I can't speak for Dave, but I'll point out that memalloc_nofs_restore is
> declared in linux/sched/mm.h, so the #include hoist makes the symbol
> available to the aops code in such a manner that now it's available to
> all the xfs code so that we don't have to remember this...

*nod*

That's historically how we've handled OS level includes - it keeps
the XFS files to including XFS headers only and that makes things
like userspace code syncs a lot easier.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com